My model suggests that the recent hype around LLMs is largely justified, but I'm updating my priors to reflect that their primary impact will be augmenting human capabilities rather than replacing them - the crux is that we're still far from true AGI, and these tools are best